Step-Audio 2

What it is

Step-Audio 2 is a speech model from the Chinese AI company StepFun (阶跃星辰). It works with sound in two ways: it can understand what an audio input contains, and it can take part in a spoken conversation. The project describes it as an end-to-end audio understanding and speech conversation model. "End-to-end" means that one model handles the whole job, taking speech in and producing a response, instead of relying on several separate programs that each handle one step.^[1]

Origins

Step-Audio 2 follows an earlier model called Step-Audio. The Step-Audio code repository now points users to Step-Audio 2.^[3] The model's technical report, which describes its design and results, is listed on arXiv, the open preprint server, under the identifier 2507.16632.^[2]

Main features

The model is built to work with several kinds of information at once. It handles semantic information, which is the meaning of the words. It also handles paralinguistic information, meaning how something is said, such as tone and pace, and non-vocal information, meaning sounds that are not speech. From the voice alone it can infer a speaker's age and emotion.^[1]

The feature set also covers tool calling. This means the model can decide to call an outside program, for example one that looks up information, and then use the result in its answer. It supports multimodal retrieval-augmented generation (RAG). In a RAG system, the model searches an outside collection of material before answering, and in Step-Audio 2 that material can include speech. One example the project gives is switching voices based on the retrieved speech.^[1]

Released versions

Step-Audio 2 comes in several versions. The mini and mini Base versions were open-sourced on 29 August 2025. A reasoning variant, Step-Audio 2 mini Think, followed on 15 September 2025. A reasoning model works through a problem step by step before it gives an answer.^[1]

The open mini models and the code are released under the Apache 2.0 license. This is a permissive open-source license: it allows people to use, modify and redistribute the software, including for commercial purposes, as long as the license notices are kept.^[1]

Reported results

The project publishes scores on several benchmarks. A benchmark is a standard test set, so that different models can be compared on the same questions. The figures below are the ones the project reports.

• On URO-Bench, a test of speech-to-speech conversation at the basic level, Step-Audio 2 scored 83.32 in Chinese and 83.90 in English.^[1]
• On MMAU, a benchmark of audio understanding, Step-Audio 2 averaged 78.0, and the mini version averaged 73.2.^[1]
• For speech recognition, the average word error rate (WER) in English was 3.14%, and the average character error rate (CER) in Chinese was 3.08%. Both measure how far a transcript differs from the correct text, so lower numbers are better.^[1]
• On Big Bench Audio, a set of audio-based questions, Step-Audio 2 mini Think scored 90.7 and Step-Audio 2 scored 80.0.^[1]

Why it matters

Step-Audio 2 shows one model covering both audio understanding and voice conversation, with tool calling and retrieval built in. Because the mini models and code are open, researchers and developers can download them and examine how the system behaves, rather than only reading a description.^[1]

Limits

The scores above come from the project's own tests on particular benchmarks. A benchmark score does not show how a model will perform in every real-world setting, such as a noisy room or a phone call with poor audio quality, and those settings need their own testing. Each benchmark also measures only part of what a model can do, so a single score should not be read as a complete measure of the model.